CDEMapper: enhancing National Institutes of Health common data element use with large language models.

Area of Science:

• Biomedical Informatics
• Data Science
• Health Research

Background:

• Common Data Elements (CDEs) are crucial for standardizing data collection and sharing across studies.
• Challenges exist in implementing CDEs due to the wide variety of data elements.
• Improved data interoperability and research reproducibility are key goals.

Purpose of the Study:

• To develop a novel Common Data Element (CDE) mapping tool to bridge the gap between local data elements and National Institutes of Health (NIH) CDEs.
• To leverage large language models (LLMs) for efficient and accurate CDE mapping.
• To enhance data interoperability in health research.

Main Methods:

• Developed CDEMapper, an LLM-powered tool with modules for CDE indexing/embeddings, CDE recommendations (combining Elasticsearch/BM25 with GPT), and human review.
• Indexed and embedded NIH CDEs for semantic search.
• Evaluated tool accuracy and usability against manual methods.

Main Results:

• CDEMapper provides a publicly available, intuitive interface for CDE mapping.
• The LLM-augmented approach (BM25 with GPT embeddings and ranker) demonstrated superior performance.
• Usability testing confirmed the tool's effectiveness and efficiency.

Conclusions:

• Large language models (LLMs) show significant potential for assisting with CDE mapping.
• The tool helps researchers identify gaps between local data and NIH CDEs.
• Promotes CDE reusability and enhances data standardization in research.